1#version 450
2
3#extension GL_KHR_shader_subgroup_shuffle: enable
4#extension GL_EXT_shader_explicit_arithmetic_types_int8: enable
5#extension GL_EXT_shader_explicit_arithmetic_types_int16: enable
6#extension GL_EXT_shader_explicit_arithmetic_types_int64: enable
7#extension GL_EXT_shader_explicit_arithmetic_types_float16: enable
8
9layout (local_size_x = 8) in;
10
11layout(binding = 0) buffer Buffers
12{
13    i8vec4 i8;
14    u8vec4 u8;
15    i16vec4 i16;
16    u16vec4 u16;
17    i64vec4 i64;
18    u64vec4 u64;
19    f16vec4 f16;
20} data[4];
21
22void main()
23{
24    uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;
25
26    data[invocation].i8.x   = subgroupShuffle(data[0].i8.x,    invocation);
27    data[invocation].i8.xy  = subgroupShuffle(data[1].i8.xy,   invocation);
28    data[invocation].i8.xyz = subgroupShuffle(data[2].i8.xyz,  invocation);
29    data[invocation].i8     = subgroupShuffle(data[3].i8,      invocation);
30
31    data[invocation].i8.x   = subgroupShuffleXor(data[0].i8.x,    invocation);
32    data[invocation].i8.xy  = subgroupShuffleXor(data[1].i8.xy,   invocation);
33    data[invocation].i8.xyz = subgroupShuffleXor(data[2].i8.xyz,  invocation);
34    data[invocation].i8     = subgroupShuffleXor(data[3].i8,      invocation);
35
36    data[invocation].u8.x   = subgroupShuffle(data[0].u8.x,    invocation);
37    data[invocation].u8.xy  = subgroupShuffle(data[1].u8.xy,   invocation);
38    data[invocation].u8.xyz = subgroupShuffle(data[2].u8.xyz,  invocation);
39    data[invocation].u8     = subgroupShuffle(data[3].u8,      invocation);
40
41    data[invocation].u8.x   = subgroupShuffleXor(data[0].u8.x,    invocation);
42    data[invocation].u8.xy  = subgroupShuffleXor(data[1].u8.xy,   invocation);
43    data[invocation].u8.xyz = subgroupShuffleXor(data[2].u8.xyz,  invocation);
44    data[invocation].u8     = subgroupShuffleXor(data[3].u8,      invocation);
45
46    data[invocation].i16.x   = subgroupShuffle(data[0].i16.x,    invocation);
47    data[invocation].i16.xy  = subgroupShuffle(data[1].i16.xy,   invocation);
48    data[invocation].i16.xyz = subgroupShuffle(data[2].i16.xyz,  invocation);
49    data[invocation].i16     = subgroupShuffle(data[3].i16,      invocation);
50
51    data[invocation].i16.x   = subgroupShuffleXor(data[0].i16.x,    invocation);
52    data[invocation].i16.xy  = subgroupShuffleXor(data[1].i16.xy,   invocation);
53    data[invocation].i16.xyz = subgroupShuffleXor(data[2].i16.xyz,  invocation);
54    data[invocation].i16     = subgroupShuffleXor(data[3].i16,      invocation);
55
56    data[invocation].u16.x   = subgroupShuffle(data[0].u16.x,    invocation);
57    data[invocation].u16.xy  = subgroupShuffle(data[1].u16.xy,   invocation);
58    data[invocation].u16.xyz = subgroupShuffle(data[2].u16.xyz,  invocation);
59    data[invocation].u16     = subgroupShuffle(data[3].u16,      invocation);
60
61    data[invocation].u16.x   = subgroupShuffleXor(data[0].u16.x,    invocation);
62    data[invocation].u16.xy  = subgroupShuffleXor(data[1].u16.xy,   invocation);
63    data[invocation].u16.xyz = subgroupShuffleXor(data[2].u16.xyz,  invocation);
64    data[invocation].u16     = subgroupShuffleXor(data[3].u16,      invocation);
65
66    data[invocation].i64.x   = subgroupShuffle(data[0].i64.x,    invocation);
67    data[invocation].i64.xy  = subgroupShuffle(data[1].i64.xy,   invocation);
68    data[invocation].i64.xyz = subgroupShuffle(data[2].i64.xyz,  invocation);
69    data[invocation].i64     = subgroupShuffle(data[3].i64,      invocation);
70
71    data[invocation].i64.x   = subgroupShuffleXor(data[0].i64.x,    invocation);
72    data[invocation].i64.xy  = subgroupShuffleXor(data[1].i64.xy,   invocation);
73    data[invocation].i64.xyz = subgroupShuffleXor(data[2].i64.xyz,  invocation);
74    data[invocation].i64     = subgroupShuffleXor(data[3].i64,      invocation);
75
76    data[invocation].u64.x   = subgroupShuffle(data[0].u64.x,    invocation);
77    data[invocation].u64.xy  = subgroupShuffle(data[1].u64.xy,   invocation);
78    data[invocation].u64.xyz = subgroupShuffle(data[2].u64.xyz,  invocation);
79    data[invocation].u64     = subgroupShuffle(data[3].u64,      invocation);
80
81    data[invocation].u64.x   = subgroupShuffleXor(data[0].u64.x,    invocation);
82    data[invocation].u64.xy  = subgroupShuffleXor(data[1].u64.xy,   invocation);
83    data[invocation].u64.xyz = subgroupShuffleXor(data[2].u64.xyz,  invocation);
84    data[invocation].u64     = subgroupShuffleXor(data[3].u64,      invocation);
85
86    data[invocation].f16.x   = subgroupShuffle(data[0].f16.x,    invocation);
87    data[invocation].f16.xy  = subgroupShuffle(data[1].f16.xy,   invocation);
88    data[invocation].f16.xyz = subgroupShuffle(data[2].f16.xyz,  invocation);
89    data[invocation].f16     = subgroupShuffle(data[3].f16,      invocation);
90
91    data[invocation].f16.x   = subgroupShuffleXor(data[0].f16.x,    invocation);
92    data[invocation].f16.xy  = subgroupShuffleXor(data[1].f16.xy,   invocation);
93    data[invocation].f16.xyz = subgroupShuffleXor(data[2].f16.xyz,  invocation);
94    data[invocation].f16     = subgroupShuffleXor(data[3].f16,      invocation);
95}
96